In recent years, the rapid advancement of generative artificial intelligence has significantly expanded the horizons of knowledge management and information retrieval. With the increasing reliance on large language models across various domains, the efficiency of context retrieval for RAG systems has become a key part of leveraging these technologies. This thesis specifically addresses the optimization of context retrieval in RAG systems, concentrating on datasets with distinct linguistic attributes such as written and spoken language. The study’s findings indicate that organizing the knowledge base according to the nature of the language (written vs. spoken) of documents enhances retrieval accuracy. Furthermore, fine-tuning source-specific embedding models proves effective in bridging performance gaps among data sources. Additionally, the research unveils promising synergies between generating alternative queries and re-ranking retrieved context using large language models, innovations primarily explored in isolation in prior literature.